School of Surveying and Spatial Information Systems

The University of New South Wales


Feature Extraction and Image Processing of Laser Scan Data

by Warren Rolfe

Supervised by B. Donnelly

Edited by J. M. RŁeger

October 2003


Aim

Feature extraction operators are used to extract features of an object from image data. Those that detect edges are known as edge operators. This thesis explored the use of three edge operators on scan data. This scan data will be obtained by scanning several objects using the CYRAX laser scanner. The scanner data can be displayed as 2-D images, from which comparisons and conclusions can be made.

Equipment

The CYRAX 2400 laser scanner was used to capture the scan data for this thesis. The scanner consists of the scanner head, a power source, a laptop computer and a tripod. Inside the scanner head are high speed rotating mirrors, which control the direction of the laser pulses. These pulses are reflected by both natural and man-made surfaces. The distance to each point is calculated using the return travel time of the signal. The horizontal and vertical angls of the mirrors are recorded. This enables three dimensional coordinates for each point to be computed. The Cyclone software is used in conjunction with the scanner to store and edit the scan data. The software can be run via the laptop that is linked to the scanner. The progress of the scan can be viewed in real-time on the laptop screen. Four walls from different buildings around the University were scanned and the resulting data edited using the Cyclone software.

Figure 1: CYRAX 2400 Laser scanner with author

Edge Detection

Three edge operators were chosen for testing, namely the Roberts , the Sobel and the Laplacian operators. One program for each edge operator was written using the Visual Basic 6.0 programming language. After the scan data was exported from Cyclone, the programs could be applied. The edges in the resulting filtered images were detected. A greyscale range from 0 to 255 was used to represent the different values of each pixel. The edges in the images can be best described as a jump in the intensity of the pixel values. In terms of an intensity profile across the scan data, an edge will be evident by a step. After the data has been filtered (see g'(x) in Fig. 2), an edge will is flagged by a large peak in the pixel values.

Figure 2: Profile Across Image Data:† (a) Edge in scan data, (b) Edge in filtered image.

Results

Figure 3 shows the results of the three edge operators being applied to one of the scans. The Sobel operator produced the clearest edges. The Laplacian was the least effective at detecting edges. The Laplacian also produced the mostnoisy images. Image noise is produced by small variations in the pixel values, not large enough to trigger an edge. The texture of the scanned surface is one reason for the noise. Although being of different quality, all three operators detected the same edges.

Figure 3: Scanned Building

             

Figure 4: Filtered Images Showing Detected Edges: Left: Laplacian operator; Middle: Roberts operator; Right: Sober operator


 

Further Information

For more information, please contact:

Brian Donnelly (Supervisor)
Email:† B.Donnelly@unsw.edu.au

Mail:
School of Surveying and Spatial Information Systems
University of New South Wales
UNSW SYDNEY  NSW  2052
Australia

Phone: +61-2-9385-4202
Fax: +61-2-9313-7493
WWW: http://www.gmat.unsw.edu.au